← Back to Sessions

Machine Learning-based Geospatial Liquefaction Model for California: A Comparative Study

Description: 

Geospatial liquefaction models commonly employ logistic regression to estimate liquefaction probability from ground shaking intensity and geospatial site parameters, valued for its straightforward probabilistic output and physical interpretability. However, these simple models may not capture complex non-linear relationships in the data, motivating the exploration of more advanced machine learning techniques. We use a newly compiled comprehensive database of liquefaction observations from 19 historic California earthquakes (Subedi et al., 2026). Using this dataset, we trained and evaluated several machine learning algorithms — Random Forests, eXtreme Gradient Boosting (XGBoost), Support Vector Machines (SVM), and Neural Networks — to predict liquefaction probability. We compared these models against the conventional logistic regression baseline model in terms of predictive accuracy, interpretability, and computational cost. Preliminary results show that the ensemble tree-based models (Random Forest and XGBoost) achieve the highest prediction accuracy, effectively capturing complex nonlinear patterns in the data and outperforming the simpler logistic regression approach. All the machine learning models demonstrate improved predictive skill over the baseline; however, their increased complexity introduces trade-offs. This comparative study highlights the trade-off between predictive performance and model interpretability in liquefaction hazard modeling. Our findings underscore the potential of modern machine learning algorithms to enhance regional liquefaction prediction accuracy in California, while also emphasizing the importance of considering model simplicity and transparency for practical engineering use.

Session: Data-Driven Advances in Liquefaction Hazard Analysis [Poster]

Type: Poster

Date: 4/17/2026

Presentation Time: 08:00 AM (local time)

Presenting Author: Hooman Shirzadi

Student Presenter: Yes

Invited Presentation: 

Poster Number: 22

Authors